KR20200075120A - Business default prediction system and operation method thereof - Google Patents

Abstract

Shown are a corporate default prediction system and an operation method thereof. According to various embodiments of the present invention, a method of predicting a default risk of a company may comprise the steps of: collecting a plurality of news articles on the Internet; selecting a company to be analyzed; classifying news articles related to an analysis target company among the plurality of collected news articles into analysis target news articles; calculating a risk level for each of the analyzed news articles; generating feature vectors representing each group by grouping the analysis target news articles based on the calculated risk level; and calculating a default risk of the analysis target company based on the generated feature vectors.

Description

Enterprise default prediction system and its operation method {BUSINESS DEFAULT PREDICTION SYSTEM AND OPERATION METHOD THEREOF}

Various embodiments of the present invention relate to a company default prediction system and a method for operating the same, and are related to a system for evaluating default risk of a company based on news data analysis.

In the modern society, the largest and most important economic actors are corporations. In a year, numerous companies are created and destroyed, and economically impacting individuals, companies, and even the country. Therefore, analyzing the rise and fall of companies is the basis not only for the industry group to which the company belongs, but also for the analysis of the entire industry group. There are various reasons for a company to disappear, but among them, if the company disappears through default, not only the employees of the company, but also the investors or trading companies of the company are greatly affected.

Accordingly, research on predicting the possibility of default for each company has been conducted steadily. In general, a company's defaultability was predicted through interest compensation ratios, debt ratios, etc., which are indicators of how much the company's operating profit can afford. As described above, financial data that can be quantitatively obtained in predicting a company's bankability analysis have been mainly used, but recently, a methodology for analyzing a company's risk based on non-financial data has emerged.

There is a possibility that the financial data disclosed by companies may not reflect the disadvantages to the company, and there are cases in which the reliability of financial reports provided by companies is also suspected. Accordingly, various methodologies for analyzing non-financial data have emerged, and news article data having an amount sufficient to be analyzed among non-financial data has been mainly used.

The amount of news article data is large enough to be analyzed, but there have been many difficulties in determining what company the news article is about and determining how to relate the news article to the risk of default. .

Korean Registered Patent No. 10-1599675

Various embodiments of the present invention aim to provide a method for predicting the default risk of a specific company based on a plurality of news articles.

Various embodiments of the present invention aim to improve the accuracy of default risk prediction through various machine learning algorithms and analysis methods.

The problems to be solved by the present invention are not limited to the problems mentioned above, and other problems not mentioned will be clearly understood by those skilled in the art from the following description.

Method for predicting the default risk of a company according to various embodiments of the present invention for solving the above-mentioned problems includes: collecting a plurality of news articles on the Internet; Selecting a company to be analyzed; Classifying news articles related to a company to be analyzed among the plurality of collected news articles into articles to be analyzed; Calculating a risk level for each of the articles to be analyzed; Generating characteristic vectors representing each group by grouping the articles to be analyzed based on the calculated risk level; And calculating the default risk of the analysis target company based on the generated characteristic vectors.

The step of calculating the risk level for each of the articles to be analyzed may include regression or itemization analysis on news articles collected by employing a specific machine learning algorithm and using the adopted machine learning algorithm. It may be to calculate a risk level for each of the articles to be analyzed by using a risk calculation algorithm derived through itemized analysis.

The company's default risk prediction method analyzes only news articles published within a certain period of time from the default of the company, among the news articles related to the default, in performing regression or itemized analysis on the collected news articles The method may further include selecting a target.

Among the collected news articles, classifying news articles related to an analysis target company into analysis target articles includes: selecting news articles including the name of the analysis target company; And determining whether each of the selected news articles is an article related to the analysis target company.

The step of determining whether each of the selected news articles is an article related to the company to be analyzed identifies a context or topic for each of the selected news articles, and the identified context or topic is associated with the company information to be analyzed. It may further include the step of determining whether or not.

The generating of characteristic vectors representing each group by performing grouping of the analysis target articles based on the calculated risk level is based on the calculated risk level for the analysis target articles, the risk level is a plurality of sections It may further include the step of categorizing as and performing grouping of the articles to be analyzed.

The step of classifying the risk level into a plurality of sections may be characterized in that classification is performed in different ways according to the type of industry to which the analysis target company belongs.

The step of calculating the default risk of the analysis target company based on the generated property vectors is to classify the property vectors of the analyzed target articles through the decision tree algorithm and perform the grouping based on the classification result. The method may further include calculating a default risk prediction value.

In the step of classifying the feature vectors of the articles to be analyzed through the decision tree algorithm, the grouping includes the feature vectors generated for the group of the articles to be analyzed into n pieces, and then the n data sets The method may further include classifying feature vectors by applying a decision tree algorithm using an n-cross validation method.

According to another embodiment of the present invention, a computing system for predicting the risk of default of a company based on news articles includes: a news article collection unit collecting a plurality of news articles on the Internet; An analysis company selection unit for selecting a company to be analyzed; An analysis article classification unit for classifying news articles related to the analysis target company among the plurality of collected news articles as analysis analysis articles; A risk level calculator that calculates a risk level for each of the articles to be analyzed; A feature vector generator for generating a feature vector representing each group by grouping the articles to be analyzed based on the calculated risk level; And a default risk prediction unit for calculating default risk of the analysis target company based on the generated characteristic vector.

According to embodiments of the present invention, for a specific company, it is possible to select news articles that are directly related to the company and to predict the default risk of the company based only on the articles.

According to the embodiments of the present invention, an effective machine learning algorithm may be employed to predict the default risk, and the risk level for each article may be calculated independently, and the prediction of default risk may be analyzed through grouping of news articles. Accuracy can be improved.

The effects of the present invention are not limited to the effects mentioned above, and other effects not mentioned will be clearly understood by those skilled in the art from the following description.

1 is a screen schematically showing an environment in which an enterprise default prediction system according to an embodiment of the present invention operates.
2 is a flowchart schematically illustrating a method of operating a corporate default prediction system according to an embodiment of the present invention.
3 is a diagram illustrating an operation in which a default prediction apparatus according to an embodiment of the present invention collects news articles.
4 is a diagram illustrating a method of classifying a news article related to an analysis target company into an analysis target article by using the NER algorithm by the default prediction apparatus according to an embodiment of the present invention.
5 is a diagram illustrating a process of selecting an independent variable used by the default prediction apparatus according to an embodiment of the present invention in calculating a risk level for articles to be analyzed.
FIG. 6 is a diagram illustrating a process in which a default prediction apparatus according to an embodiment of the present invention selects a machine learning algorithm to be used in calculating risk levels for articles to be analyzed.
7 is a diagram illustrating a process in which a default prediction apparatus according to an embodiment of the present invention predicts a default of a company based on risk levels of articles to be analyzed.
8 is a diagram illustrating a method of grouping articles to be analyzed related to a specific company by the default prediction apparatus according to an embodiment of the present invention.
9 is a diagram for explaining a method of calculating a default risk of an analysis target company through the classified characteristic vectors according to the default prediction apparatus according to an embodiment of the present invention.
FIG. 10 is a diagram illustrating a cross-validation method used by the default prediction apparatus according to an embodiment of the present invention to classify feature vectors through a decision tree algorithm.
11 is a block diagram showing the configuration of a default prediction apparatus according to an embodiment of the present invention.
12 is a flowchart illustrating a process for a default prediction apparatus according to an embodiment of the present invention to derive a default risk prediction value of an analysis target company.

The terminology used herein is for describing the embodiments and is not intended to limit the present invention. In the present specification, the singular form also includes the plural form unless otherwise specified in the phrase. As used herein, “comprises” and/or “comprising” does not exclude the presence or addition of one or more other components other than the components mentioned. Throughout the specification, the same reference numerals refer to the same components, and “and/or” includes each and every combination of one or more of the components mentioned. Although "first", "second", etc. are used to describe various components, it goes without saying that these components are not limited by these terms. These terms are only used to distinguish one component from another component. Therefore, it goes without saying that the first component mentioned below may be the second component within the technical spirit of the present invention.

When a certain part of the specification "includes" a certain component, it means that the component may be further included other than excluding other components, unless otherwise specified. In addition, terms such as "... unit" and "module" described in the specification mean a unit that processes at least one function or operation, which may be implemented in hardware or software, or a combination of hardware and software. .

1 is a screen schematically showing an environment in which an enterprise default prediction system according to an embodiment of the present invention operates.

Referring to FIG. 1, the corporate default prediction system may include a default prediction apparatus 100, a user terminal 200, and an external server 300.

The default predicting apparatus 100 may collect default news articles and predict default risk for a specific company through analysis of the collected news stories.

The default prediction apparatus 100 may be configured as a computing system having memory means and a computing device. That is, the default prediction apparatus 100 may be configured as a server having an intensive processing function. Alternatively, a memory means such as a personal computer (for example, a desktop computer, a notebook computer, etc.), a workstation, a PDA, a web pad, etc. It may be configured with one of digital devices equipped with a computing power equipped with a microprocessor. In the memory means included in the default prediction apparatus 100, software that implements functions related to default prediction may be stored or installed.

The user terminal 200 may be a terminal used by a user who wants to obtain information related to the default risk of a specific company through communication with the default prediction apparatus 100.

The user terminal 200 according to an embodiment of the present invention is a mobile phone, a smart phone, a PDA (Personal Digital Assistant), a PMP (Portable Multimedia Player), a tablet PC, etc., which can be connected to a web server through a network. It may include a handheld-based wireless communication device, a personal computer (for example, a desktop computer, a notebook computer, etc.), a workstation, a PDA, a web pad, etc., equipped with a memory means and equipped with a microprocessor It could be one of the digital devices with computational power.

According to one embodiment, the user transmits identification information on the company that wants to know the risk of default using the user terminal 200 to the default prediction device 100, and the corresponding company analyzed by the default prediction device 100 Information about the default risk prediction may be received and confirmed through the user terminal 200.

The external server 300 is a server that is not managed by the management entity of the default prediction apparatus 100, and according to an embodiment, the external server 300 may be a server including news data. The default prediction apparatus 100 may access the external server 300 to collect news data. The collection of news data performed by the default prediction apparatus 100 may be performed by crawling various types of news data published on the web. In this case, the default prediction apparatus 100 may collect news data from a plurality of external servers 300. In this way, the number of external servers 300 is not limited to a specific number.

The default prediction apparatus 100 may perform communication with the user terminal 200 and the external servers 300 through a communication network implemented in various ways.

The communication network may be implemented as one of a wired communication network, a wireless communication network, and a composite communication network. For example, the communication network may include a mobile communication network such as 3G, Long Term Evolution (LTE), and LTE-A. The communication network may include a wired or wireless communication network such as Wi-Fi, Universal Mobile Telecommunizations System (UMTS)/General Packet Radio Service (GPRS), or Ethernet. The communication networks are Magnetic Secure Transmission (MST), Radio Frequency IDentification (RFID), Near Field Communication (NFC), ZigBee, Z-Wave, Bluetooth, Low Power Bluetooth (BLE, Bluetooth Low Energy) Or, it may include a local area network, such as infrared communication (IR, InfraRed communication). The communication network may include a local area network (LAN), a metropolitan area network (MAN), or a wide area network (WAN). 2 is a flowchart schematically illustrating a method of operating a corporate default prediction system according to an embodiment of the present invention.

Referring to FIG. 2, the default predicting apparatus 100 may roughly perform a three-step operation to finally predict the default risk of a company.

Step (a) of FIG. 2 is a step of performing pre-processing by collecting news data, step (b) is a step of calculating a risk level for each news article, and step (c) is finally performing a company default risk prediction It is a step.

Referring to step (a) of FIG. 2, in step (a), the default prediction apparatus 100 may access the Internet to collect news articles through crawling. In this process, the default prediction apparatus 100 may collect only news articles that include the name of the analysis target company selected as the analysis target.

Subsequently, the default prediction apparatus 100 may determine whether the collected news articles are related to the company to be analyzed and classify a part of the plurality of news articles as the analysis target articles. For example, even if there is a news article containing the name of the company to be analyzed in the form of text, there may be a case where the name is used for a purpose other than the name of the company to be analyzed on the text of the news article. have. Also, there may be cases where the context or the subject of the article is not related to the company to be analyzed. In the above cases, the news article includes the name of the company to be analyzed, but the articles may not be classified as the article to be analyzed.

Referring to step (b) of FIG. 2, in step (b), the default predicting apparatus 100 may calculate a risk level for each of the analysis targeted articles classified as being related to the company to be analyzed.

In step (b), risk level calculation for each of the articles to be analyzed may be performed through machine learning. That is, the default prediction apparatus 100 sets at least some of the collected news articles as training data for machine learning and test set, and adopts a specific machine learning algorithm and then applies the corresponding machine learning algorithm. Through this, it is possible to derive a risk calculation algorithm used to calculate the risk level by performing analysis on learning data and applying it to test data to evaluate the analysis results. Subsequently, the default prediction apparatus 100 may calculate the risk level for the articles to be analyzed through the derived risk calculation algorithm.

The default prediction apparatus 100 may adopt a machine learning algorithm having the best evaluation result by performing learning and testing on various types of machine learning algorithms in the above manner and evaluating the result. The adoption of the machine learning algorithm according to the evaluation result may be automatically performed by the default prediction apparatus 100, or the administrator of the default prediction apparatus 100 may check and select the evaluation result.

According to an embodiment, the learning data includes information on which company has caused bankruptcy, and the default prediction apparatus 100 may use such information to evaluate the machine learning algorithm. For example, the evaluation of machine learning algorithms may be performed according to the degree to which the risk level of a news article for a company that has defaulted is calculated high and the risk level of a news article for a company that has not defaulted to low is calculated.

In step (b), a risk calculation algorithm derived by the finally adopted machine learning algorithm is utilized, so that a risk level for each of the articles to be analyzed can be calculated.

Referring to step (c) of FIG. 2, the default predicting apparatus 100 may finally perform a default risk prediction of a company based on the risk level of news articles calculated in step (b).

The default predicting apparatus 100 may calculate the default risk of the target company in a numerical form based on the level of risk of the target articles analyzed for the target company. In addition, the default predicting apparatus may determine the default of the default for the company based on the calculated default risk, and may additionally calculate the reliability of the calculated default risk.

The default predicting apparatus 100 performs grouping of the analysis target articles based on the calculated risk level and generates characteristic vectors for each group in calculating the default risk of the analysis target company based on the risk levels of the analysis target articles. Therefore, the risk of default can be calculated based on this.

3 is a diagram illustrating an operation in which the default prediction apparatus 100 according to an embodiment of the present invention collects news articles.

Referring to FIG. 3, the default prediction apparatus 100 may collect news articles through crawling on the Internet. 3, only news articles for a specific company are displayed among news articles collected by the default prediction apparatus 100.

According to an embodiment, the default prediction apparatus 100 may select only a news article including a specific company name from among all news articles using a morphological analysis library such as Lucene.

4 is a diagram illustrating a method of classifying a news article related to an analysis target company into an analysis target article using the NER algorithm by the default prediction apparatus 100 according to an embodiment of the present invention.

Referring to FIG. 4, the default prediction apparatus 100 may utilize a specialized Entity Recognition (NER) algorithm in classifying articles related to a company to be analyzed. As described above in FIG. 3, only news articles including the name of a specific company among all news articles may be selected by utilizing a morpheme analysis library, etc., but in this case, the selected articles are classified as articles to be analyzed and used. Several problems existed below. Looking at the problems, there were cases where certain news was not related to the company at all, and there were cases in which morphological analysis library allowed articles that were judged to be slightly misspelled, so that unrelated articles were selected. In addition, there were cases where articles with different names were selected, such as the name of a specific company not only used as the name of the company, but also used as the name of the product sold by the company.

In order to solve the above problems, the default predicting apparatus 100 utilizes the NER algorithm to determine whether the news articles selected through the morpheme analysis library are news related to the company to be analyzed and the corresponding articles as the analysis target articles Can be classified. That is, the default prediction apparatus 100 selects news articles including the name of the company to be analyzed using a morpheme analysis library, and for each of the news articles selected through the NER algorithm, the article is related to the company to be analyzed. You can judge whether it is an article.

According to an embodiment, the NER algorithm may be implemented through a programming language R, and may be implemented using POS tagging and n-gram methods. POS (Part-Of-Speech) tagging refers to a method of classifying text according to a grammatical function or form, and such classification can be performed through various criteria. The n-gram may mean a method of classifying sentences into groups having n syllables or words and performing analysis on the words in a grouping manner.

Referring to FIG. 4, it may be determined whether articles selected for a company to be analyzed through the NER algorithm are related to the company to be analyzed. Based on whether or not the text containing the name of the company has been extracted, it is possible to determine whether the article is related to the company to be analyzed by applying the NER algorithm, and as shown in FIG. 4,'ACCEPT' or'REJECT' and Results can be derived in the same form of yes or no.

As described above, according to the NER algorithm, the context or subject for each of the selected news articles is identified, and whether the identified context or subject is related to the company information to be analyzed can be determined, and the name of the specific company is the corresponding company It may be determined whether or not it is used in a meaning other than.

FIG. 5 is a diagram illustrating a process of selecting an independent variable used by the default prediction apparatus 100 according to an embodiment of the present invention in calculating the risk level for the articles to be analyzed.

The default prediction apparatus 100 may utilize a machine learning algorithm when performing analysis on the articles to be analyzed classified through the above-described method through FIGS. 3 and 4. The default predicting apparatus 100 may derive a risk calculation algorithm by adopting a specific machine learning algorithm, setting some of the articles to be analyzed as training data, and then performing analysis on it. The machine learning algorithm used in this process can use regression analysis or itemized analysis.

Referring to FIG. 5, the default prediction apparatus 100 may select independent variables to be used for calculating the risk level through various methods. The independent variables used to calculate the risk level may include at least some of the sentences or words included in news articles.

According to an embodiment, the default prediction apparatus 100 may classify and select words included in news articles in an n-gram format, and may select words through various feature selection methods. In addition, a word to be used as an independent variable in performing risk level calculation may be selected based on the scarcity calculated according to the frequency with which words appear in a news article.

5 shows the results when independent variables are selected in various ways. In the graph on the right side of FIG. 5, the horizontal axis represents a method in which independent variable selection is made differently, the bar graph represents the number of independent variables used in each method, and the line graph represents the evaluation score for each method.

Referring to Figure 5, as can be seen in the 1-3 and 1-4 method, as the number of independent variables increases, the evaluation score may increase accordingly, but in this method, an overfitting problem may also occur. The way that it handles the number of independent variables that is too large can overwhelm the computing system's resources. Therefore, as shown in the 3-1 method and the 3-2 method, if an appropriate number of independent variables is selected, an appropriate evaluation score can be calculated accordingly without overfitting.

FIG. 6 is a diagram illustrating a process in which the default prediction apparatus 100 according to an embodiment of the present invention selects a machine learning algorithm to be used for calculating a risk level for articles to be analyzed.

Referring to FIG. 6, the default prediction apparatus 100 may adopt various types of machine learning algorithms, analyze articles to be analyzed, and derive a risk calculation algorithm based on the analysis results. According to one embodiment, the machine learning algorithm may use regression analysis or itemized analysis. The type of the machine learning algorithm may include MLP Regression, Logistic Regression, Decision Tree, Random Forest, Adaboost Classifier, SVM (Support Vector Machine), and the like as illustrated in FIG. 6. However, the types of machine learning algorithms that the default prediction apparatus 100 can adopt and use are not limited to the above examples.

FIG. 6 shows numerical values obtained by evaluating a result of analyzing an article to be analyzed through various types of machine learning algorithms through AUC and Gini Value. Area Under Curve (AUC) is a calculation method used in statistics to evaluate the performance of discriminant models. It means the area below. The maximum value of AUC is 1, and the higher the value, the better the performance of classification. The Gini coefficient is also one of the calculation methods used in statistics to evaluate the performance of discriminant models.The area between the ROC curve in the graph and the diagonal line starting from the origin and classifying the graph in half is called A, and the area under the ROC curve is When B is called, it can be obtained by dividing A by (A+B). The relationship between the AUC and the Gini coefficient may be a relationship in which the Gini coefficient is calculated by subtracting 1 from the value obtained by multiplying AUC by 2.

The default predicting apparatus 100 may calculate a risk level for each of the articles to be analyzed by adopting a specific machine learning algorithm based on the above evaluation results and using a risk calculation algorithm derived through the algorithm.

7 is a diagram illustrating a process in which the default predicting apparatus 100 according to an embodiment of the present invention predicts default of a company based on risk levels of articles to be analyzed.

Referring to FIG. 7, the default predicting apparatus 100 may determine the ratio of default articles based on the level of risk of the target articles of analysis related to the analysis target company, and may calculate default risk accordingly. Referring to FIG. 7, there is an approximate error, but it is indicated that the proportion of default articles gradually increases as time progresses. (DD: Distance to Default) is also shown to increase. It can be seen from the portion 701 indicated by a circle in FIG. 7 that the ratio of default articles temporarily increases, but the default risk value does not show a large change. As such, the default predicting apparatus 100 may control the default risk value to not change even if the default rate is temporarily increased or decreased.

8 is a diagram illustrating a method in which the default prediction apparatus 100 according to an embodiment of the present invention groups analysis target articles related to a specific company.

Referring to FIG. 8, risk levels for a plurality of articles to be analyzed are represented in a numerical form, and a form in which grouping of articles to be analyzed is performed based on a risk level is displayed.

According to an embodiment of the present invention, a plurality of articles to be analyzed may be displayed in various levels of risk levels as shown in FIG. 8, which may be grouped into a plurality of groups according to any criteria based on the risk level. In FIG. 8, the analysis target articles for which the risk levels of 0.7, 0.8, and 0.9 were calculated were grouped into the first group 801, and the analysis target articles for which the risk levels of 1.2 and 1.3 were calculated were grouped for the second group 803. , 0.3, 0.4, and 0.5 were calculated, and the analysis targeted articles were grouped into the third group 805. As described above, according to an embodiment of the present invention, a plurality of sections according to a risk level are set, and articles to be analyzed belong to each section, so grouping of the articles to be analyzed may be performed.

9 is a diagram for explaining a method of calculating a default risk of an analysis target company through the classified characteristic vectors according to the default prediction apparatus 100 according to an embodiment of the present invention.

Referring to FIG. 9, the default prediction apparatus 100 may generate feature vectors representing each group after the analysis target articles are grouped based on the risk level. The characteristic vector may include a statistical value calculated based on the risk level of the analyzed articles belonging to each group as an element of the vector. Statistics calculated based on the level of risk of the articles to be analyzed may include a minimum value (Minimum), a maximum value (Maximum), an average (Mean), a median (Median), a mode (mode), and the like.

After generating the characteristic vectors for each group, the default prediction apparatus 100 may calculate the default risk of the analysis target company based on the generated characteristic vectors. According to an embodiment, the default predicting apparatus 100 may calculate default risk and finally determine the default risk of the company in a yes or no form.

In this process, the default prediction apparatus 100 may classify feature vectors through a decision tree algorithm. That is, the feature vectors can be classified based on various types of features of the feature vectors, and based on the classification result, the predicted default risk of the company to be analyzed can be calculated, and the reliability of the predicted default risk can also be calculated.

According to an embodiment, the default prediction apparatus 100 may use the probability distribution value of each class in calculating the default risk prediction value through a decision tree algorithm. For example, the default risk prediction value can be calculated based on the probability that the same class will appear within the same branch of the decision tree, that is, the characteristic vectors used for data classification are classified into a specific class.

The default prediction apparatus 100 may also calculate the reliability of the default risk prediction in addition to the default risk prediction, which decreases the reliability when the default risk prediction value is calculated without reviewing all the properties of the feature vectors by the decision tree algorithm. It can be calculated in the form. For example, if there are 10 criteria in which feature vectors can be classified on a decision tree algorithm, and the final class classification is completed while only a part of the criteria are used, and the default risk prediction value is calculated, the number of criteria used is small. Reliability calculation can be performed in such a way that the reliability decreases as it is calculated.

FIG. 10 is a diagram illustrating a cross-validation method used by the default prediction apparatus 100 according to an embodiment of the present invention to classify feature vectors through a decision tree algorithm.

Referring to FIG. 10, when the default prediction apparatus 100 classifies feature vectors through a decision tree algorithm, classification may be performed through an n-cross validation method. In detail, the default prediction apparatus 100 classifies the data into n data based on data including risk level information calculated for the articles to be analyzed, and then sets one of the n data sets to test data ( Test set), the remaining set can be repeated n times using the training set (Training set or Learning set). As described above, according to the cross-validation method, the data set used as test data in the iterative process is continuously changed, and analysis can be repeated up to n times with different test data and learning data.

11 is a block diagram showing the configuration of the default prediction apparatus 100 according to an embodiment of the present invention.

Referring to FIG. 11, the default prediction apparatus 100 includes a news article collection unit 110, an analysis company selection unit 120, an analysis target article classification unit 130, a risk level calculation unit 140, and a characteristic vector generation unit It may be configured to include 150, the default risk prediction unit 160, the communication unit 170, the storage unit 180 and the control unit 190.

For convenience of description, the subjects performing each role in the default prediction apparatus 100 are indicated in the form of ~ units, but each of the parts is a subprogram module or control unit operating in the default prediction apparatus 100 ( 190) may be functionally classified. These program modules include, but are not limited to, routines, subroutines, programs, objects, components, data structures, etc. that perform each operation or execute a specific abstract data type.

The news article collection unit 110 may collect news articles on the Internet. To this end, the news article collection unit 110 may access various types of external servers 300. The news article collection unit 110 may collect news articles in a manner of crawling news-related data published on the web.

According to an embodiment, the news article collection unit 110 may classify and collect only articles related to a corresponding company when a company to be analyzed is selected. In performing such a function, the news article collection unit 110 may use the functions of the analysis company selection unit 120 and the analysis target article classification unit 130, which will be described later.

The analysis company selection unit 120 may select a company to be analyzed. As described above, the news article collection unit 110 may collect only articles related to the company selected by the analysis company selection unit 120, and the analysis target article classification unit 130, which will be described later, also analyzes the company selection unit ( 120) can categorize articles about the companies they choose. According to an embodiment, the analysis company selection unit 120 may select a plurality of companies to be analyzed, or may select a company corresponding to company information received from the user terminal 200 as a company to be analyzed.

The analysis target article classification unit 130 may classify news articles related to an analysis target company among the collected news articles as analysis target articles. To this end, the article classification unit 130 to be analyzed may primarily select news articles including the name of the company to be analyzed and determine whether each of the selected news articles is an article related to the company to be analyzed. .

The article classification unit 130 to be analyzed may utilize a morphological analysis library in selecting a news article including the name of the company to be analyzed. In addition, the analysis target article classification unit 130 identifies the context or topic for each of the selected news articles when the news articles including the name of the analysis target company are selected, and the identified context or subject is analyzed with the company information to be analyzed. You can determine whether there is a connection. For example, if the company to be analyzed is a toy manufacturing business, and the name of the company is included and one of the selected articles contains semiconductor related content that is not related to this, the article classification unit 130 to be analyzed The article can be judged to be unrelated to the company information being analyzed and can be excluded from the article being analyzed.

The risk level calculating unit 140 may calculate a risk level for each of the analysis target articles classified by the analysis target article classification unit 130. In this process, the risk level calculator 140 may adopt a specific machine learning algorithm and perform regression analysis or itemized analysis on news articles collected using the machine learning algorithm. The risk level calculator 140 may calculate a risk level for each of the analyzed articles using a risk calculation algorithm derived as a result of regression analysis or itemized analysis.

The risk level calculating unit 140 may perform analysis on all news articles collected by the news article collection unit 110 as well as news about the company to be analyzed. That is, since the risk level calculator 140 needs to learn the corresponding machine learning algorithm when using the machine learning algorithm, news articles such as these may be used as training data. According to an embodiment, the risk level calculating unit 140 may manage not only news articles but also corporate default information. The default information of the company may include information related to the time when the default occurred and the company information that caused the default.

In performing analysis on news articles collected through a machine learning algorithm, the risk level calculator 140 may target only news articles about companies that are bankrupt, and news about companies that are bankrupt. For articles, only news articles published within a certain period of time from the company's default may be selected for analysis. For example, the risk level calculator 140 may analyze only news articles issued for two years before the default date when the default occurs.

As described above, the risk level calculating unit 140 analyzes only news articles issued within a predetermined time before the company defaults, through the machine learning algorithm, and uses the risk calculation algorithm derived according to the analysis results. By calculating the risk level for each of the articles to be analyzed, the accuracy of the risk level calculation can be improved. That is, the risk level calculating unit 140 performs learning through articles of companies that are aware that bankruptcy has occurred in the step of performing learning of the machine learning algorithm, and performs the risk calculation algorithm derived through the machine learning algorithm. Using it, the risk level can be calculated for each news article of the analysis target company that has not occurred until now.

The feature vector generator 150 may group the articles to be analyzed based on the risk level calculated for each article to be analyzed and generate feature vectors representing each group.

The feature vector generator 150 classifies the analysis target articles into a plurality of groups according to the section of the risk level calculated for the analysis target articles, and sets the group based on the statistical value of the risk level of the analysis target articles included in each group. It is possible to generate representative feature vectors.

According to an embodiment, the risk level statistic of the analysis target articles used by the feature vector generator 150 to generate the feature vectors is the mean (Mean), median (Median), mode (Mode), and minimum value of the risk level values. (Minimum), the maximum value (Maximum), and the like. The feature vector generator 150 may generate a feature vector using these statistics or other values calculated using the same as an element of the feature vector.

According to an embodiment, when classifying the risk level into a plurality of sections, the characteristic vector generator 150 may classify the risk level in different ways according to the type of industry to which the analysis target company belongs. For example, the characteristic vector generator 150 classifies industries into manufacturing, medical, financial, and telecommunications industries, identifies where the company to be analyzed belongs to, and then performs risk level section classification according to each industry type. Can. For another example, the feature vector generator 150 may classify the industry to which the analysis target company belongs as a manufacturing industry or a non-manufacturing industry, and perform the risk level section classification accordingly.

The default risk predicting unit 160 may calculate a default risk predicted value of a corresponding company based on a plurality of characteristic vectors generated for a company to be analyzed, and calculate reliability associated with calculating the predicted value.

According to an embodiment, the default risk prediction unit 160 may calculate default risk prediction values by classifying feature vectors through a decision tree algorithm. The default risk predicting unit 160 classifies data including feature vectors generated for a group of articles to be analyzed into n pieces, and then applies a decision tree to n data sets in an n-cross validation method. Classification of feature vectors can be performed.

The default risk prediction unit 160 may calculate a default risk prediction value based on a probability distribution value of each class on a decision tree. In addition, in calculating the reliability of the default risk prediction, the default risk prediction unit 160 calculates the reliability of the default risk prediction based on the number of characteristics reviewed by the decision tree algorithm in the process of calculating the default risk prediction value. Can.

The communication unit 170 allows the default prediction apparatus 100 to communicate with the user terminal 200 and the external server 300. The communication network used by the communication unit 170 to perform communication may be configured regardless of its communication mode such as wired or wireless, and may be implemented as various types of communication networks.

The storage unit 180 may serve to store data utilized in the process of operating the default prediction apparatus 100. For example, the storage unit 180 may store and manage collected news article data and analysis data therefor.

The storage unit 180 may include, for example, a memory, a cache, a buffer, and the like, and may be composed of software, firmware, hardware, or a combination of at least two of them. According to an embodiment, the storage unit 180 may be implemented in the form of a read only memory (ROM).

The control unit 190 includes a news article collection unit 110, an analysis company selection unit 120, an analysis article classification unit 130, a risk level calculation unit 140, a characteristic vector generation unit 150, and a default risk prediction unit 160, a function of controlling data flow between the communication unit 170 and the storage unit 180 may be performed. That is, the control unit 190 according to the present invention includes a news article collection unit 110, an analysis company selection unit 120, an analysis target article classification unit 130, a risk level calculation unit 140, and a characteristic vector generation unit 150 ), the default risk prediction unit 160, the communication unit 170 and the storage unit 180 can be controlled to perform a unique function, respectively.

In FIG. 11, a news article collection unit 110, an analysis company selection unit 120, an analysis target article classification unit 130, a risk level calculation unit 140, a characteristic vector generation unit 150, and a default risk prediction unit 160 ) Is a functional classification of the control unit 190, and thus may be integrated into one control unit 190.

12 is a flowchart illustrating a process for the default predicting apparatus 100 according to an embodiment of the present invention to derive a predicted default risk value of an analysis target company.

Referring to FIG. 12, the default prediction apparatus 100 may collect news articles on the Internet (S1201). The news article data accumulated in this collection step may be used as training data or test data of a machine learning algorithm in the future, and only news articles related to a company selected as a company to be analyzed may be separately classified and used for predicting default risk.

The bankruptcy prediction apparatus 100 may perform news article analysis through a machine learning algorithm and derive a risk calculation algorithm (S1203). In this process, the default prediction apparatus 100 may adopt a specific machine learning algorithm by performing news article analysis through various types of machine learning algorithms and comparing the results. The bankruptcy prediction apparatus 100 may collect and manage bankruptcy information of companies separately from news article data, and may evaluate news article analysis results based on this.

The bankruptcy prediction apparatus 100 may select a company to be analyzed and classify news articles related to the company as an analysis target article (S1205). The selection of the company to be analyzed may be performed by the user and the selection information may be transmitted from the user terminal 200 to the default prediction apparatus 100, or may be performed by the manager of the default prediction apparatus 100. . The bankruptcy prediction apparatus 100 may classify news articles related to a corresponding company among news articles collected in advance, but may also search and collect news articles related to the relevant company on the Internet. The default prediction apparatus 100 may collect only articles including the name of the company to be analyzed by utilizing the morpheme library.

The bankruptcy prediction apparatus 100 selects a company to be analyzed and classifies news articles related to the company into articles to be analyzed, and includes the name of the company to be analyzed among all news articles collected using the morpheme library. Only news articles can be selected, and only news articles related to the company to be analyzed among the selected articles using the NER algorithm can be classified as articles to be analyzed. In contrast, the default prediction apparatus 100 classifies only news articles related to a company to be analyzed as articles to be analyzed by using the NER algorithm when only news articles including the name of the company to be analyzed are collected in the collecting step. can do.

After classifying the analysis target article, the default prediction apparatus 100 may calculate the risk level for each of the analysis target articles using the risk calculation algorithm derived in step S1203 (S1207).

When the risk level for each of the articles to be analyzed is calculated, the default prediction apparatus 100 may perform grouping of the articles to be analyzed based on the calculated risk level (S1209). According to an embodiment, grouping of articles to be analyzed may be performed based on which of the plurality of sections the risk level calculated for each article belongs to.

The default prediction apparatus 100 may generate characteristic vectors for each group of articles to be analyzed by using a result of grouping the articles to be analyzed (S1211 ). The elements included in the feature vectors may be various statistics calculated based on the risk level of the articles to be analyzed included in each group, or values calculated using these statistics.

The default predicting apparatus 100 may finally calculate the default risk of the analysis target company based on the generated feature vectors (S1213). Calculation of default risk of the company to be analyzed can be performed through the process of classifying feature vectors through a decision tree algorithm. In addition, the default prediction apparatus 100 may separately calculate the reliability of the default risk prediction in the process of calculating the default risk prediction value.

As described above, according to various embodiments of the present invention, it is possible to calculate the default risk of a company through analysis of news articles about a specific company, and the bank can perform a credit evaluation of companies based on this, and Loan risk management can be strengthened. In addition, an algorithm suitable for calculating default risk may be selected among various machine learning algorithms, and the level of risk through each news article may be calculated independently, and the overall default risk may be predicted through grouping of news articles.

The steps of a method or algorithm described in connection with an embodiment of the present invention may be implemented directly in hardware, a software module executed by hardware, or a combination thereof. Software modules include random access memory (RAM), read only memory (ROM), erasable programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), flash memory, hard disk, removable disk, CD-ROM, or It may reside on any type of computer readable recording medium well known in the art.

The embodiments of the present invention have been described above with reference to the accompanying drawings, but a person skilled in the art to which the present invention pertains may implement the present invention in other specific forms without changing its technical spirit or essential features. You will understand. Therefore, it should be understood that the above-described embodiments are illustrative in all respects and not restrictive.

100: default prediction device
200: user terminal
300: external server

Claims (10)

A method for predicting a company's default risk based on news articles, performed by a computing system, comprising:
Collecting a plurality of news articles on the Internet;
Selecting a company to be analyzed;
Classifying news articles related to a company to be analyzed among the plurality of collected news articles into articles to be analyzed;
Calculating a risk level for each of the articles to be analyzed;
Generating characteristic vectors representing each group by grouping the articles to be analyzed based on the calculated risk level; And
And calculating the default risk of the analysis target company based on the generated characteristic vectors.
According to claim 1,
The step of calculating the risk level for each of the articles to be analyzed,
After adopting a specific machine learning algorithm and performing a regression or itemization analysis on news articles collected using the adopted machine learning algorithm, the analysis target is performed using a risk calculation algorithm derived through the regression or itemization analysis A method for predicting a company's default risk by calculating the risk level for each of the articles.
According to claim 2,
In performing a regression or itemized analysis on the collected news articles, further comprising the step of selecting only the news articles published within a certain time from the default of the company among the news articles related to the default company How to predict the default risk of a company.
According to claim 1,
The step of classifying the news articles related to the analysis target company into the analysis target articles among the plurality of collected news articles,
Selecting news articles containing the name of the company to be analyzed; And
Further comprising the step of determining whether each of the selected news articles related to the analysis target company, the method for predicting the default risk of the company.
According to claim 4,
For each of the selected news articles, determining whether the article is related to the analysis target company,
And identifying a context or subject for each of the selected news articles, and determining whether the identified context or subject is related to the analyzed company information.
According to claim 1,
The step of generating the characteristic vectors representing each group by grouping the articles to be analyzed based on the calculated risk level,
Based on the risk level calculated for the articles to be analyzed, further comprising the step of classifying the risk level into a plurality of sections and performing grouping of the articles to be analyzed through this, a method for predicting default risk of a company.
The method of claim 6,
The step of classifying the risk level into a plurality of sections,
Method for predicting default risk of a company, characterized in that the classification is performed in different ways according to the type of industry to which the analysis target company belongs.
According to claim 1,
Calculating the default risk of the analysis target company based on the generated characteristic vectors,
A method of predicting the default risk of a company, further comprising classifying characteristic vectors of the articles to be analyzed, which have been grouped through a decision tree algorithm, and calculating the default risk prediction value of the analyzed target company based on the classification result.
The method of claim 8,
The step of classifying the feature vectors of the articles to be analyzed by grouping through the decision tree algorithm,
After classifying the data including the feature vectors generated for the group of articles to be analyzed into n, and applying a decision tree algorithm to the n data sets in a cross-validation method (n-Cross Validation) A method for predicting default risk of a company, further comprising classifying the vectors.
In a computing system that predicts a company's default risk based on news stories,
A news article collection unit collecting a plurality of news articles on the Internet;
An analysis company selection unit for selecting a company to be analyzed;
An analysis article classification unit for classifying news articles related to the analysis target company into the analysis target articles among the collected plurality of news articles;
A risk level calculator that calculates a risk level for each of the articles to be analyzed;
A feature vector generator for generating a feature vector representing each group by grouping the articles to be analyzed based on the calculated risk level; And
And a default risk prediction unit for calculating default risk of the analysis target company based on the generated characteristic vector.

Images